Integritetsbevarande falldetektering med federerad maskininlärning i IoT-baserade applikationer

Björn Dahlström () och Axel Svegerud ()

Sammanfattning

For use cases involving the sharing of personal and sensitive data, maintaining user trust is crucial. When individuals feel confident that their privacy and data integrity are protected, they are more likely to participate, resulting in a larger and more diverse dataset. This, in turn, leads to the development of higher quality models and products that better serve customer needs.

Moreover, the implementation of privacy-preserving techniques like federated learning offers significant cost benefits. By training and running models directly on users' devices, we can reduce the need for cloud resources, ultimately lowering operational costs while still delivering valuable insights. This approach not only enhances privacy but also makes the system more efficient and scalable.

The project will:

(1) implement ML models that detect falls, using an open dataset.
(2) investigate how federated learning can be applied to the use case of fall detection. Ideally we would want to be able to run models on the user’s phone. Realistically we start by running instances of the platform on computers that
can send the training results to a mother model.
(3) enable data sharing between different user groups and apply use of differential privacy. Since data collected over the phone will be extensive and sensitive, the project team could enable data sharing in the platform using privacy-preserving technology.